NVIDIA KGMON Engineers Reliable Data Analytics Agent Harness for KDD Cup

The NVIDIA KGMON team secured second place in the KDD Cup 2026 Data Agents competition, delivering a framework that demonstrates how reliable autonomous analytics depend on system design rather than model scale. Tasked with resolving natural-language queries across heterogeneous inputs including relational databases, CSV and JSON files, prose documents, and briefing videos, participants were constrained to a fixed small language model. This limitation forced teams to optimize the agent harness itself. KGMON’s placement highlights a critical industry pivot: deterministic performance in data workflows is achieved through constrained execution environments, not open-ended generative capability. The winning architecture centralizes data access by converting all structured inputs into a unified SQLite interface. A preflight schema-scanning step automatically identifies join keys, unit mismatches, and null distributions, feeding this context to the agent before reasoning begins. This eliminates routing errors and preserves inference capacity for analytical logic. The team further restricted the model through an opinionated tool harness that limits operations to a minimal set of verified functions. A persistent Python environment maintains state between calls and includes middleware to automatically repair malformed requests, preventing single-point failures from aborting entire runs. By centralizing queries, document lookups, and answer generation, the framework drastically reduces syntax errors and wasted turns. For unstructured inputs, the system applies targeted preprocessing rather than raw model processing. Policy texts and reports are scanned using character-limited previews and regex matching, with relevant excerpts routed to a separate zero-temperature extraction layer. This keeps the primary agent context window focused on computation. Similarly, briefing videos are converted into timestamp-aligned transcripts and keyframes before the agent loop, avoiding prohibitive computational overhead during execution. Reliability is enforced through rigorous observability and iterative refinement. Every attempt generates detailed execution traces, logging prompts, tool calls, intermediate outputs, and error states. This transparency enables both automated inspector agents and human reviewers to categorize failures, distinguish schema confusion from formatting errors, and prioritize system updates. The team leverages controlled ensembling, grouping results by answer values and deploying additional attempts only when confidence thresholds are unmet, thereby balancing reliability gains against compute costs. Crucially, the framework emphasizes disciplined governance over autonomous escalation. While the system supports iterative improvement loops for prompt and tool tuning, strict review gates, held-out validation tasks, and human sign-off prevent benchmark overfitting. Human engineers retain oversight of success criteria and reusable skill integration, while agents handle execution and trace analysis. As enterprises scale autonomous data workflows, the KDD Cup results signal a shift toward modular, harness-centric agent design. By enforcing constrained action spaces, normalizing data interfaces, and embedding continuous trace evaluation, developers can deploy smaller, cost-effective models that deliver auditable, repeatable analytics.